feat(qy): QY ops backend with FA and longctx MoE/RoPE kernels

[Ceng23333]

Summary

• Add QY (Denglin) native CUDA ops backend with linked FlashAttention (varlen + KV-cache) paths.
• Add longctx MoE/RoPE QY kernels needed by MiniCPM5 FA/longctx/cudagraph stack (ITW release/2026.10.13-qy-minicpm5-16a3-fa-longctx-cudagraph).
• Rebased onto current master; clang-format applied to touched C++ files.

Validation

• Smoke build / pytest -m smoke (QY): not run in this environment (no QY/Denglin SDK on the PR host). Please maintainers with QY access run:

  python -m pip install .[dev] --no-build-isolation --no-deps \
    --config-settings=cmake.define.INFINI_RT_ROOT=/path/to/infini-rt-prefix \
    --config-settings=cmake.define.INFINI_OPS_SMOKE_BUILD=ON
  python -m pytest tests -m smoke -q

• Full suite deferred pending QY platform access.
• Integration-validated via InfiniTensorWorktree pin release/2026.10.13-qy-minicpm5-16a3-fa-longctx-cudagraph (MiniCPM5 longctx + CUDA graph capture fixes).

Related

• Companion: InfiniRT QY device backend PR (same stack)
• Downstream: InfiniCore submodule pin + InfiniLM MiniCPM5 FA/longctx PR onto refactor/adopt-modern-infini-stack

Made with Cursor

Stack PR index (merge order)

1. InfiniOps: feat(qy): QY ops backend with FA and longctx MoE/RoPE kernels #1002
2. InfiniRT: feat(qy): enable QY CUDA runtime device backend InfiniRT#49
3. InfiniCCL: feat(qy): QY-compatible CPU device header fix InfiniCCL#77
4. InfiniCore: chore(qy): pin InfiniRT/Ops/CCL for MiniCPM5 FA/longctx InfiniCore#1575
5. InfiniLM: feat(qy): MiniCPM5 FA, longctx MoE/RoPE, CUDA-graph capture safety InfiniLM#597 → base refactor/adopt-modern-infini-stack

ITW pin: https://github.com/Ceng23333/InfiniTensorWorktree/tree/release/2026.10.13-qy-minicpm5-16a3-fa-longctx-cudagraph

[Ceng23333]

QY required smoke (PR head)

• Head SHA: ed0b40db75fc5fdb7e638401fb6bba657b372857
• Host: Denglin QY, DENGLIN_DEVICES=3
• Command: pytest tests/test_add.py tests/test_gemm.py --devices qy (m0 ops config build against InfiniRT PR head)
• Result: PASS (1633 passed, 3 skipped)
• Log: …/logs/prhead-verify/ops-pytest-add-gemm.log

1633 passed, 3 skipped in 18.51s

[Ceng23333]

QY required smoke (PR head) — detailed

• Head SHA: ed0b40db75fc5fdb7e638401fb6bba657b372857
• Host: Denglin QY, DENGLIN_DEVICES=3
• Build: m0 ops config (add,gemm) vs InfiniRT PR head 95d30fe
• Command: pytest tests/test_add.py tests/test_gemm.py --devices qy
• Overall: PASS
• Log: …/logs/prhead-verify/ops-pytest-add-gemm.log (+ per-file ops-pytest-per-file.log)

Suite	Device	Passed	Skipped	Failed	Wall time
tests/test_add.py	qy	133	3	0	1.32s
tests/test_gemm.py	qy	1500	0	0	15.70s
Total (M0 bar)	qy	1633	3	0	18.51s

Combined run: 1633 passed, 3 skipped in 18.51s.